Skip to content

The Threat Landscape of Vibe Coding and the Path to Safe Autonomy

Threat Landscape of Vibe Coding and the Road to Secure Autonomy

arXiv: 其他 AI 安全论文· arXiv:2611.21822· Devki Nandan Jha, Shubh Anand, Karthick Keeranagere Krishnaiah, Roman Ahearn, Suhas Nale·Indexed by this site · 2 days ago· Originally publishedDate not givenOpinion· Its place in the scan record →
AI summary

Researchers propose a security threat landscape for vibe coding, integrating research scattered across prompt injection, insecure code generation, supply chain attacks, and human factors into a unified threat model. The paper establishes a reference architecture, characterizes six levels of agent autonomy, and builds a threat model around five trust boundaries: the developer, the AI model, tools, cloud infrastructure, and the broader software ecosystem, proposing a taxonomy of nine major categories comprising 21 threats, with evidence, evaluated defenses, and limitations laid out for each. The authors' analysis concludes that risk rises as agents' operational permissions expand, while most defenses rely on model judgment and address only isolated threats. The paper closes with 11 research challenges, including enforceable limits on agent permissions and shared benchmarks and methods for evaluating vibe coding security.

Read originalarxiv.org